{"id":"eed54260-c96b-473d-8e5e-6299f0a44e24","arxiv_id":"2505.19389","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors release MIMICEL, an event log of emergency department patient journeys derived from MIMIC-IV-ED, complete with quality validation and usage examples.","lead":"This paper builds a public process-mining event log, MIMICEL, from the MIMIC-IV-ED emergency department database, covering 425,028 ED visits and 7.6 million events. It validates the log and demonstrates how it supports length-of-stay and crowding analyses that could help hospitals improve emergency care flow.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The synthetic triage timestamp (intime+1s, Methods Step 7) makes 'Triage in the ED' a deterministic relabeling of entry time; the claim that 'no errors were introduced' is untenable and triage-to-discharge statistics in Tables 9-10 are likely artifacts.","rationale":"The reader's weakest assumption correctly identifies the synthetic triage timestamp as the most load-bearing premise. My stress-test sharpens that concern: it is not merely that triage time is a proxy, but that the +1s rule makes 'Triage in the ED' a deterministic function of entry time, so any interval involving triage is a transformation of entry/discharge times, not an independent measurement. The paper's Technical Validation claim that 'no errors were introduced during the log extraction process' is directly contradicted by the introduction of a synthetic event, even though the construction is disclosed in Step 7. The acuity-based analysis in Tables 9-10 reports 'Triage to Discharge' as a distinct pathway, which is misleading because the activity is invented. Despite this, the dataset has substantial value: the other five activities use real timestamps, the extraction scripts are public, and the validation surfaces genuine data quality issues. The paper already discloses the triage construction, so the appropriate outcome remains conditional acceptance with a required revision to explicitly recommend against using triage as an observed activity, or to remove triage from the core event log. The concrete test I propose would settle whether the concern lands by demonstrating the tautological nature of the triage-based statistics.","tokens_in":13995,"tokens_out":4326,"duration_ms":37092,"concrete_test":"For each case counted in the 'Triage to Discharge' path of Table 9, verify two facts: (1) the activity sequence is exactly Enter → Triage → Discharge with no intervening events, and (2) the reported 'Triage to Discharge' interval equals outtime - intime - 1s. If both hold for every such case, the triage activity is an artifact of the +1s rule, and Tables 9-10 are LoS-based statistics rather than triage-based statistics. A complementary check is to rerun the acuity analysis with the triage timestamp set to intime + 15 minutes instead of +1s; if the 'Triage to Discharge' frequencies and intervals change materially, the published numbers are highly sensitive to an arbitrary modeling choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Methods Step 7 assigns 'Triage in the ED' an artificial timestamp equal to intime + 1 second because the MIMIC-IV-ED triage table lacks timestamps. This is not a harmless approximation: it forces triage to be the second event in every case, and for any case following the 'Triage to Discharge' path, the reported interval is exactly outtime - intime - 1s, i.e., the length of stay minus one second. Consequently, the 'Triage to Discharge' path frequency and interval in Tables 9-10 and the corresponding arrow in Figure 4 do not describe observed triage-to-discharge behavior; they describe a subset of cases where no other recorded activity occurs between entry and discharge. The Technical Validation section concludes 'no errors were introduced during the log extraction process,' yet a synthetic event type was introduced and then treated as a real activity in the demonstration analyses. The paper discloses the construction, but it does not quantify how sensitive the acuity-based insights are to this choice. This matters because one of the six activity types is not an observed event, and all triage-related process maps and derived performance metrics inherit this artifact. The central claim that MIMICEL is a high-fidelity, end-to-end event log is therefore weaker than stated: the log is useful, but triage-based analyses should be relabeled or removed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes the curation of MIMICEL, an event log extracted from the MIMIC-IV-ED dataset, following the nine-step guideline of Jans et al. The log records six activity types for 425,028 ED stays, is provided in CSV and XES formats on PhysioNet, and is accompanied by publicly available extraction and validation code. The authors validate the log with the DaQAPO framework and present three demonstration analyses (acuity-based, length-of-stay, and crowdedness) to illustrate the log's utility for ED process mining. The central claim is that MIMICEL captures an end-to-end patient journey with high fidelity and that no errors were introduced during extraction.","tokens_in":14389,"tokens_out":4622,"duration_ms":43805,"significance":"If the concerns below are addressed, MIMICEL would fill a genuine gap: it provides a large, publicly available, reproducible event log for ED process mining and makes MIMIC-IV-ED accessible to the process-mining community. The authors follow a systematic extraction guideline, publish their SQL and Python scripts, and ship the log in two standard formats, which are concrete strengths. However, the paper's central validation claim is currently overstated because one activity is synthetic and because duplicate discharge events are treated as faithful records. The dataset remains valuable, but its description and the demonstrated analyses need to be made more cautious and precise.","major_comments":[{"comment":"The 'Triage in the ED' activity is assigned the synthetic timestamp intime + 1 second because the MIMIC-IV-ED triage table has no timestamp. This makes triage a deterministic relabeling of 'Enter the ED' rather than an observed event. Consequently, the 'Triage to Discharge' path in Figure 4 and Tables 9-10 can occur only for cases with no other recorded activity between entry and discharge, and the reported interval is exactly outtime - intime - 1 second (i.e., the length of stay minus one second). The paper discloses this construction, but it does not assess how the acuity-based conclusions are affected; for example, the statement that 'lower-acuity patients are more likely to be discharged directly after triage' is an interpretation of an artifact. Please either remove triage-based analyses, explicitly relabel them as 'entry-to-discharge with no intermediate recorded activity,' or provide evidence that intime+1 is a clinically valid proxy for triage time.","section":"Step 7, Table 2, Figure 4, Tables 9-10"},{"comment":"The statement that 'no errors were introduced during the log extraction process' is contradicted by two extraction choices. First, the synthetic triage event is introduced by the authors, not inherited from MIMIC-IV-ED. Second, converting each diagnosis row into a separate 'Discharge from the ED' event causes 60.13% of cases to have multiple discharge events at identical timestamps; this is an extraction-induced representation, not a source-data issue. These choices inflate the total event count (7,568,824) and can distort process-discovery results, including self-loops and path frequencies. Please revise the validation claim to distinguish source-data issues from extraction-induced representations, and either aggregate discharge events per stay or publish a variant of the log with one discharge event per case.","section":"Technical Validation, 'Remark' and 'Repeated activities with identical timestamps'"},{"comment":"The two main demonstration analyses rely on threshold-like choices that are derived from the same dataset: the 500-minute LoS threshold is the 75th percentile of the data, and the crowdedness threshold of 12 simultaneous patients is the 75th percentile of the simultaneous-patient distribution. The paper presents these as descriptive cohort comparisons, which is acceptable, but it should explicitly state that the thresholds are data-derived and that the resulting 'normal' versus 'prolonged' or 'crowded' versus 'non-crowded' categories are relative to this cohort rather than externally validated. Otherwise readers may over-interpret the quantitative claims in Tables 9-12.","section":"Usage Notes, 'Acuity-based analysis' and 'LoS-driven analysis'"}],"minor_comments":[{"comment":"The years in the timestamp column (2165, 2169, 2170) appear to be typos; please replace them with the actual years present in MIMIC-IV-ED, and correct 'W ALK IN' to 'WALK IN'.","section":"Table 6"},{"comment":"In the bullet list, 'In contrast.' should be 'In contrast,' and in the fourth bullet the phrase '50 and 36 minutes (Table 10' is missing a closing parenthesis.","section":"Usage Notes, 'Acuity-based analysis'"},{"comment":"The text says the incomplete_cases function was used to detect incomplete ED visits, but no result for this check is reported in Table 7 or in the narrative; please add the result or remove the mention.","section":"Technical Validation"},{"comment":"The claim that 500 minutes 'aligns with the internationally recommended acceptable ED LoS of ≤8 hours' is not quite right: 8 hours is 480 minutes. Please correct the comparison or use a threshold of 480 minutes if the intent is to align with the cited recommendation.","section":"Usage Notes, 'LoS-driven analysis'"},{"comment":"The process map is dense and the thick red arrow may be difficult to read in a black-and-white print version; please increase font size or add a zoomed inset for the triage-to-discharge path.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the data-descriptor scope and the reproduction package is a real asset. My main concern is that the Technical Validation section overstates fidelity: the synthetic triage event and the duplicate discharge events are extraction choices, not merely inherited source-data issues. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection. If the authors add a caution in the Usage Notes about the triage proxy and provide a single-discharge-event variant of the log, the central claim would become defensible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real contribution—a public, code-backed event log from MIMIC-IV-ED—but the synthetic triage timestamp is a load-bearing simplification. It makes the paper's 'no errors were introduced' claim untenable and turns the triage-to-discharge statistics in Tables 9 and 10 into arithmetic on length of stay. Worth reviewing, but it needs revision before anyone builds on the triage-related numbers.\n\nWhat's new and good: this is the first systematic application of the Jans et al. guideline to MIMIC-IV-ED. The nine-step extraction is documented in detail, code is on GitHub, and the log is on PhysioNet. Validation with DaQAPO is appropriate and surfaces real issues—15.1% of HOME cases with a hadm_id, pain values outside 0-10, and so on. The duplicated discharge events and repeated med reconciliation at identical timestamps are disclosed and explained by attribute granularity; that's defensible.\n\nThe soft spots are real. The triage activity is always intime+1 second. So the 'Triage to Discharge' path in Table 9 and the interval in Table 10 describe only cases where no vitals, meds, or other recorded activity happened between entry and discharge, and the interval is literally outtime minus intime minus one second. That is not observed triage behavior; it's a proxy. The paper discloses the construction but never quantifies how much the acuity-based insights depend on it. The Technical Validation's closing remark—'no errors were introduced during the log extraction process'—contradicts the fact that a new event type was invented. That sentence should be rewritten. Also, the exact MIMIC-IV-ED version needs to be specified: the reference says v2.2 but the PhysioNet log page says 2.1.0. Readers regenerating the log need to know which source matches.\n\nMinor: the 500-minute and 12-patient thresholds are data-driven, but they're only used to split cohorts for descriptive comparison, not to fit a model, so that's not circular in a damaging way. The 1.64% missing acuity is disclosed.\n\nThis is a dataset paper, not a methods paper. The audience is process mining and ED operations researchers who want a standard benchmark. It deserves a serious referee and likely conditional acceptance after the triage issue is handled—either by relabeling triage as 'arrival plus one second' in all analyses, removing triage-based claims, or adding sensitivity analysis. Fix that, and the artifact is solid.","headline":"MIMICEL is a genuinely useful public event log with code and validation, but the synthetic triage timestamp makes the 'no errors introduced' claim untenable and turns the triage-to-discharge statistics into arithmetic on length of stay.","tokens_in":14824,"tokens_out":2442,"would_cite":true,"duration_ms":21363,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A public event log now maps 425,000 emergency-department stays, extracted from MIMIC-IV-ED data and ready for process mining.","keywords":["MIMIC-IV-ED","event log","process mining","emergency department","patient flow","data quality","XES","length of stay"],"falsifier":"Obtain real triage times for a sample of visits, either from a future version of the source database that records triage timestamps or from a chart audit, and compare them with the synthetic entry-time-plus-one-second values; if the gap is more than a few minutes, the paper's triage-to-discharge intervals and any process map containing triage as a distinct step would need revision.","tokens_in":13770,"feed_emoji":"🏥","tokens_out":5731,"duration_ms":49691,"temperature":0.7,"pith_summary":"The paper's central claim is that MIMICEL, an event log curated from the public MIMIC-IV-ED emergency-department database, captures the end-to-end patient journey through the ED well enough for process-mining analysis. It reports that extraction followed a systematic nine-step guideline, yielding 425,028 ED stays with 7,568,824 events across six activity types, and that technical validation found no extraction-introduced errors: the quality problems detected were inherited from the source data. If this is right, process-mining researchers get a ready-made, reproducible dataset for studying patient flow, acuity-driven care intensity, length of stay, and crowding, without rebuilding extraction pipelines. The paper's demonstration analyses show that higher-acuity patients receive more frequent and tightly timed vital-sign and medication cycles, while lower-acuity patients are more often discharged directly after triage.","feed_headline":"New event log maps 425,000 emergency-department stays","feed_subtitle":"Extracted from public MIMIC-IV-ED data, MIMICEL lets process-mining tools analyze patient flow, acuity, and crowding directly.","key_machinery":"The machinery that carries the argument is the mapping from the throughput phase of the ED-crowding conceptual model to the MIMIC-IV-ED tables, which fixes the process boundaries and six activity types, together with the nine-step extraction guideline that determines case IDs, timestamps, attributes, and invalid-case filtering. The pivotal mechanism is the temporal ordering of each stay's events into a sequence, with the invented triage event inserted one second after entry. That ordering is what turns static relational snapshots into an event log suitable for process discovery and performance analysis. A secondary mechanism is the quality-assessment framework used to certify that the extraction did not introduce errors.","core_discovery":"The core discovery is a publicly reusable event log derived from the MIMIC-IV-ED relational tables. The authors map six activities—Enter the ED, Triage in the ED, Vital sign check, Medicine reconciliation, Medicine dispensation, and Discharge from the ED—onto the source tables, using each stay's intime and outtime as the entry and discharge boundaries, chart times from the vital-sign and medication tables, and one synthetic timestamp: triage is set to entry time plus one second because the triage table has no timestamps. Each stay_id is a case and all other source fields become case or event attributes. Validation under a systematic quality framework reports missing acuity values in 1.64% of cases, 15.1% of home-discharge cases retaining a hospital admission ID, 29% of pain values outside the 0-10 range, and mixed timestamp granularities and temperature units; the paper attributes all of these to the source dataset. Demonstration analyses then reveal monotonic acuity trends, such as consecutive vital-sign-check intervals rising from 30 minutes at acuity 1 to 120 minutes at acuity 5, and a triage-to-discharge path that becomes more common and faster as acuity drops.","pith_inferences":["The synthetic triage timestamp makes the triage activity a proxy; researchers should test whether process maps change when triage is removed or given a realistic delay.","The 15.1% of home discharges that retain a hospital admission ID could corrupt any downstream linkage between ED events and inpatient data, so users should clean disposition-dependent attributes before joining.","Because medication and vital-sign timestamps are minute-granular while entry and discharge are second-granular, sub-minute durations are unobservable and overlap-based crowding estimates are approximate.","MIMICEL could serve as a benchmark corpus for process-mining algorithm evaluation, since its ground truth is the public source data and its quality issues are documented and inherited."],"forward_implications":["Researchers can apply process-mining discovery, conformance, and performance-analysis techniques directly to MIMICEL without re-extracting from the relational source.","The acuity analyses imply a measurable gradient in care intensity: from acuity 1 to acuity 5, the median interval between consecutive vital-sign checks rises from 30 to 120 minutes.","The triage-to-discharge direct path rises from 1.23% of acuity-1 cases to 18% of acuity-5 cases, with median duration falling from 126 to 60 minutes.","For high-acuity patients with prolonged length of stay, vital-sign self-loops occur in 88% of cases and transitions between medication and monitoring take about twice as long as in fast-tracked urgent cases.","Crowded ED conditions are associated with a higher proportion of admitted patients: about one-third of admitted patients were treated in crowded conditions versus one-fourth of patients discharged home."],"supporting_citations":[{"why":"Supplies the source MIMIC-IV-ED relational dataset from which all events and attributes are extracted.","marker":"[11]"},{"why":"Provides the nine-step guided process used to structure the event-log extraction.","marker":"[12]"},{"why":"Defines the throughput model of ED crowding that determines the core activities and process boundaries.","marker":"[17]"},{"why":"Provides the emergency-department data-quality assessment framework used for validation.","marker":"[21]"},{"why":"Supplies the taxonomy that categorises the data-quality issues reported in the validation.","marker":"[22]"},{"why":"Implements the quality-assessment tool used to detect missing values, dependency violations, anomalies, and range errors.","marker":"[23]"},{"why":"Contributes the question-driven methodology and expert questions that motivated the acuity and length-of-stay analyses.","marker":"[14]"},{"why":"Provides the Python library used to convert the CSV event log to XES format.","marker":"[20]"}],"fun_headline_variants":["MIMICEL event log maps 425,000 emergency stays","New ED event log enables process mining on MIMIC-IV","MIMICEL: reusable event log for emergency department analysis","Event log from MIMIC-IV-ED tracks six patient-flow steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything that involves triage timing rests on the assumption that the triage event happened exactly one second after the patient entered the ED, because the source data records no triage timestamp.","fun_headline_variants_meta":{"raw":{"variants":["MIMICEL event log maps 425,000 emergency stays","New ED event log enables process mining on MIMIC-IV","MIMICEL: reusable event log for emergency department analysis","Event log from MIMIC-IV-ED tracks six patient-flow steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1293,"prompt_tokens":992,"completion_tokens":301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":608,"tokens_out":301,"duration_ms":3369,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:13:56.927568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Obtain real triage times for a sample of visits, either from a future version of the source database that records triage timestamps or from a chart audit, and compare them with the synthetic entry-time-plus-one-second values; if the gap is more than a few minutes, the paper's triage-to-discharge intervals and any process map containing triage as a distinct step would need revision.","supporting_citations":[{"cited_title":"& Jouck, T","cited_arxiv_id":null,"evidence_quote":"Provides the nine-step guided process used to structure the event-log extraction."},{"cited_title":"R.et al.A conceptual model of emergency department crowding.Annals emergency medicine42, 173–180 (2003)","cited_arxiv_id":null,"evidence_quote":"Defines the throughput model of ED crowding that determines the core activities and process boundaries."},{"cited_title":"& Braekers, K","cited_arxiv_id":null,"evidence_quote":"Provides the emergency-department data-quality assessment framework used for validation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the taxonomy that categorises the data-quality issues reported in the validation."},{"cited_title":"& Janssenswillen, G","cited_arxiv_id":null,"evidence_quote":"Implements the quality-assessment tool used to detect missing values, dependency violations, anomalies, and range errors."},{"cited_title":"Sci.7, 302 (2017)","cited_arxiv_id":null,"evidence_quote":"Contributes the question-driven methodology and expert questions that motivated the acuity and length-of-stay analyses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Python library used to convert the CSV event log to XES format."}],"review_version":1}